Claude Fable 5.1 cuts AI costs: what it means for your software

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, and the detail that matters most to any company building something with AI isn’t the model itself, it’s the price: cache-read costs drop 75%, from $1.00 to $0.25 per million tokens, while every other rate stays the same as Fable 5 (Anthropic’s announcement; VentureBeat). Anthropic says this brings total bills down 25% to 45% for typical and agentic workloads. At Evicron, an AI and custom software studio based in Barcelona, the first question clients ask when news like this drops is whether they should switch providers. That’s almost never the right question.

What Anthropic actually shipped

Fable 5.1 and Mythos 5.1 are, per Anthropic, the same underlying model with different levels of safeguards. Fable 5.1 is the generally available version, with the company’s standard protections in place. Mythos 5.1 is only offered through restricted-access programs, aimed at vetted cybersecurity and life-sciences organizations that need capabilities the usual safeguards constrain. It’s the same pattern we’ve seen with other frontier model releases: splitting the general-purpose product from the one that operates with fewer guardrails, under the vendor’s direct oversight.

The price cut applies specifically to cached context reads — the technique that lets a system reuse, at reduced cost, information it already processed in an earlier call (system instructions, reference documents, a long conversation’s history). The more an application depends on reusing that context — agents that hold memory across a long task, assistants that query the same knowledge base over and over — the more the savings actually show up.

Why this matters even if you don’t track tokens

Most companies that come to us for a custom software project with AI built in don’t manage token consumption directly — their technology partner does. But the price per token is the baseline that decides whether a use case pencils out at all. A support assistant that re-reads the same product manual on every reply, or an agent that reviews contracts while holding the full document in context for the whole session, are exactly the scenarios where a 75% cheaper cache changes the math: what didn’t justify automating a year ago might justify it today.

This isn’t an isolated event. We’ve spent the last year and a half watching the major providers — Anthropic, OpenAI, Google — compete on price as much as on capability. We covered this dynamic when Anthropic pulled ahead of OpenAI in the enterprise segment, and the practical takeaway hasn’t changed: don’t build your software assuming today’s provider and today’s price will still hold in a year.

What doesn’t change: where a project’s cost actually lives

Token pricing is the most visible part of an AI project’s cost, but it’s almost never the largest. Our guide to custom software development costs breaks this down in detail: solution design, integrating with the systems a company already runs, preparing the data the model will work with, and testing before launch usually outweigh the monthly API bill. A vendor’s price cut improves the margin or the business case for a well-designed project — it doesn’t turn a poorly-designed one into a good one.

What to do if you have an AI project underway or on the roadmap

1. Don’t migrate on price alone. Switching models carries testing and validation cost. A 25-45% saving on the cache portion only justifies a change if your application genuinely reuses a lot of context.
2. Revisit your AI architecture every few months. What didn’t pencil out in 2025 might today, simply because every serious provider’s per-token cost has come down.
3. Don’t lock your product to a single model. If your integration lets you switch providers without rewriting business logic, you capture these price drops without betting everything on one vendor winning the race.
4. Get an audit before you decide. Working out whether a price cut justifies a migration or a new use case takes real usage and token-volume data — not a rough guess.

How we approach this at Evicron

In our applied AI projects we work with several providers — Claude, GPT, Gemini — and pick the model based on the use case and actual cost, not out of habit. When the question is whether it’s worth revisiting the AI architecture you already have, or whether a use case that didn’t add up before now does, that’s where our AI consulting for businesses comes in: we audit real usage, compare providers against your own data, and tell you, with numbers, whether moving something is worth it.

Bottom line

Anthropic’s price cut on Claude Fable 5.1 confirms that the cost of building with AI keeps falling, and the advantage goes to whoever designs their software to capture those drops without getting locked into one vendor. Before deciding whether it affects you, measure how much context your use case actually reuses — that’s where the answer is.

Have an AI project underway and want to know if it’s worth revisiting at today’s prices? Get in touch: the first consultation is free and we reply within 24 hours.